onnx / onnx/optimizer

[BUG] The pass "eliminate_nop_cast" causes output mismatch

Open
#273 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
C++
Stars
834
Forks
109
Avg merge
6h 55m
Merged PRs (30d)
2

Description

[BUG] The pass "eliminate_nop_cast" causes output mismatch

Issue
Running the single pass eliminate_nop_cast with onnxoptimizer 0.4.2 changes numerical outputs.ORT graph optimizations are disabled during differential testing.

Environment

  • Ubuntu20.04
  • Python 3.10
  • onnx==1.19.0
  • onnxruntime==1.23.2
  • onnxoptimizer==0.4.2 (latest)

Repro steps (run from this folder)

  1. Create a Python environment (Python 3.10) and install dependencies:
python3 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -r requirements.txt
  1. Optimize a case with only eliminate_nop_cast (writes model.opt.onnx next to model.onnx):
  • python optimize_model.py --case ./case_04319_seed54453180
  • python optimize_model.py --case ./case_06878_seed74717901
  1. Differential test for original and optimized models with same oracle inputs:
  • python diff_test.py --case ./case_04319_seed54453180
  • python diff_test.py --case ./case_06878_seed74717901

Observed results
Mismatches appear in the two cases abouve.

Expected
eliminate_nop_cast should be semantics-preserving. Applying this pass alone should not change any output values.

Differential Test Output Details

  1. case_04319_seed54453180 — output[6] max_abs=3.174e-01 max_rel=1.693e-01; output[8] max_abs=4.484e+18 max_rel=4.484e+26; overall mismatch.
  2. case_06878_seed74717901 — output[4] max_abs=6.900e+01 max_rel=5.694e+01 (largest); output[1] max_abs=7.354e+00 max_rel=1.316e+00; overall mismatch.
Case: case_04319_seed54453180 (ORT graph optimizations disabled)
  output[0]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(9, 1)
  output[1]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 53)
  output[2]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(9, 53)
  output[3]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(9, 53)
  output[4]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(9, 53)
  output[5]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(9, 53)
  output[6]: max_abs=3.174e-01, max_rel=1.693e-01, shape=(9, 53)
  output[7]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(53, 9)
  output[8]: max_abs=4.484e+18, max_rel=4.484e+26, shape=(1, 1, 1)
  output[9]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(9,)
  output[10]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(9, 53)
  output[11]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(9, 53)
  output[12]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(9, 53)
  output[13]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(9, 53)
  output[14]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(9, 53)
Overall: max_abs=4.484e+18, max_rel=4.484e+26
Case: case_06878_seed74717901 (ORT graph optimizations disabled)
  output[0]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(46, 47)
  output[1]: max_abs=7.354e+00, max_rel=1.316e+00, shape=(47, 47)
  output[2]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(47, 47)
  output[3]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(47,)
  output[4]: max_abs=6.900e+01, max_rel=5.694e+01, shape=(1, 47, 49)
  output[5]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 47, 1)
  output[6]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(47,)
  output[7]: max_abs=0.000e+00, max_rel=0.000e+00, shape=()
  output[8]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(47, 47)
Overall: max_abs=6.900e+01, max_rel=5.694e+01

Attachments

  • requirements.txt
  • optimize_model.py
  • diff_test.py (runs original and optimized with same oracle inputs; ORT graph optimizations disabled)
  • run_optimized_model.py (executes optimized model with oracle inputs)
  • case_04319_seed54453180/ (original and optimized model with oracle inputs)
  • case_06878_seed74717901/ (original and optimized model with oracle inputs)

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Run optimize_model.py and diff_test.py for case_04319_seed54453180 and case_06878_seed74717901 using the supplied requirements.txt, with ORT graph optimizations disabled. Inspect the eliminate_nop_cast pass entry point and compare the original and optimized outputs; done means both cases produce identical outputs when that pass is applied alone.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp, python
Domain
machine-learning, tooling
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
38/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.